Procedia Computer Science 24 ( 2013 ) 143 – 151

1877-0509 © 2013 The Authors. Published by Elsevier B.V.Selection and peer-review under responsibility of the Program Committee of IES2013doi: 10.1016/j.procs.2013.10.037

ScienceDirect

17th Asia Pacific Symposium on Intelligent and Evolutionary Systems, IES2013

Towards Believable Resource Gathering Behaviours in Real-time

Strategy Games with a Memetic Ant Colony System

Xianshun Chena, Yew Soon Onga,∗, Liang Fenga, Meng Hiot Limb, Caishun Chena,Choon Sing Hoa

aMulti-plAtform Game Innovation Centre (MAGIC), School of Computer Engineering, Nanyang Technological University, SingaporebSchool of Electrical and Electronic Engineering, Nanyang Technological University, Singapore

Abstract

In this paper, the resource gathering problem in real-time strategy (RTS) games, is modeled as a path-finding problem where game

agents responsible for gathering resources, also known as harvesters, are only equipped with the knowledge of its immediate sur-

roundings and must gather knowledge about the dynamics of the navigation graph that it resides on by sharing information and

cooperating with other agents in the game environment. This paper proposed the conceptual modeling of a memetic ant colony

system (MACS) for believable resource gathering in RTS games. In the proposed MACS, the harvester’s path-finding and resource

gathering knowledge captured are extracted and represented as memes, which are internally encoded as state transition rules (mem-

otype), and externally expressed as ant pheromone on the graph edge (sociotype). Through the inter-play between the memetic

evolution and ant colony, harvesters as memetic automatons spawned from an ant colony are able to acquire increasing level of

capability in exploring complex dynamic game environment and gathering resources in an adaptive manner, producing consistent

and impressive resource gathering behaviors.

c© 2013 The Authors. Published by Elsevier B.V.

Selection and peer-review under responsibility of the Program Committee of IES2013.

Keywords: Memetic Computing; Ant Colony; Resource Gathering; Path Finding

1. Introduction

With the advancements in graphics and physics engines reaching the stage of maturity, game researchers are now

turning to artificial intelligence (AI) as an increasingly influential factor for enhancing the quality of user experiences

and furthering the growth of game innovation and believability. At the same time, due to the dynamic nature and

complexity of digital games, many challenges in the field have posed as interesting problems that can be addressed

using artificial intelligence and computational intelligence (CI) technologies. One of these challenges is the dynamic

path-finding and resource gathering problem commonly found in real-time strategy (RTS) games. In RTS games,

the primary objective of the game is to build an army strong enough to destroy the opponent base without getting

∗ Corresponding author. Tel.: +65-6790-6448 ; fax: +65-6792-6559.

E-mail address: asysong@ntu.edu.sg

Available online at www.sciencedirect.com

© 2013 The Authors. Published by Elsevier B.V.Selection and peer-review under responsibility of the Program Committee of IES2013

Open access under CC BY-NC-ND license.

Open access under CC BY-NC-ND license.

144 Xianshun Chen et al. / Procedia Computer Science 24 ( 2013 ) 143 – 151

overrun by the enemies. In order to do that, players are supposed to build and strengthen their base by purchasing

various upgrades and soldiers using the available game resources provided. Efficiency in gathering these scarce and

limited resources is often one of the key factors that determines the eventual outcome of the game, especially when

the location of these game resources are not known beforehand since they are often are randomly generated. This can

be a challenging task for AI if the computer controlled players, often known as a bots, were to play on equal grounds

with the human players, meaning to say that the bots are not equipped with complete information on the locations

of the resources. Rather, they have to explore the map extensively in order to locate the resources. While players

have come to expect increasing levels of AI sophistication in the resource gathering and path-finding behaviors, the

modern commercial game environments for RTS games are becoming frequently non-deterministic with imperfect

knowledge and a large number of rather dynamic variables. Furthermore, the dynamic path-finding and resource

gathering behaviors in these games have to deal with the vast number of game agents frequently as well as the time

synchronization among them in a large-scale persistent game world with the use of intensive computational resources.

Consequently, dynamic path-finding and intelligent gathering of arbitrarily valued resources in RTS games represent

a fundamental problem of high complexity, and with potential merits if appropriately addressed1.

RTS games with particularly innovative AI for dynamic path-finding and resource gathering can stand out among

competitors. Such innovative AI is related to the general principle of believability. As discussed in2, believable agent

is described as one that provides the illusion of life, and thus permits the audience’s suspension of disbelief. A be-

lievable agent is usually characterized with human-like peculiarities, such as the capabilities to learn, to cooperate,

to make mistakes, to adjust its own strategy in response to the opponent’s actions. In this work, the RTS resource

gathering problem to be solved here is first defined as a path-finding problem, while game agents that are responsible

for gathering the resources shall be referred to as harvesters in this paper. It is assumed that harvesters have very

limited knowledge of the digital game world, apart from its immediate or local surroundings. In contrast to the A-

Star search, which requires complete information of the navigation graph and essentially cheats to know the resource

location ahead of time, the harvester bots must now explore and discover resource locations by sharing information

and cooperating with other harvesters in its surroundings. This clearly puts the harvester bots in more equal footings

with the human players in the game play, hence providing greater believability in the resultant RTS game. Achieving

cooperative behavior is often difficult, as maintaining shared information about the dynamics of agents in the game

world can be complex. One approach is via explicit planning and search3,4, which has been used to tackle the cooper-

ative path-finding problem that require multiple agents to follow non-interfering paths from the current states to their

respective goal states. The approach is however computationally expensive, though well-suited for handling complex

navigation problems where movement is constrained. An alternative approach is based on swarm intelligence and

the flow-field techniques. Among such an approach is the direction map proposed in5,6, which presents a distributed

approach whereby agents share information about the direction in which they traveled when passing through each

checkpoint. The information on each checkpoint then serves to encourage agents that pass through the same location

to travel in the same direction as past agents. The Ant Colony Optimization (ACO) proposed in1 is another example

on the use of swarm intelligence for RTS resource gathering. The distributed nature of swarm intelligence is well

suited for maintaining shared information about the dynamics of agents in the game world. Nevertheless, the ap-

proach is slow in converging to the best path-finding solution as they are poor in efficiently concentrating around the

best path-finding solutions. As a result, harvesters may be spending lots of time wandering and looking silly before

they concentrate on the optimal paths1. One other limitation of such computational intelligence method is that its

performance is highly sensitive to parameter tuning and hence leading to undesirable path-finding behaviors when

the navigation graph is highly dynamic in nature, which is not acceptable to game developers. Furthermore, the cur-

rent swarm-based cooperative path-finding approach including5,6,1 are not designed for game environments that are

based on navigation mesh7, which have become the search space representation of choice for path-finding in digital

games8,9.

In this paper, our interest is on the social behavior of harvester bots in RTS resource gathering tasks. In particular,

we propose a Memetic Ant Colony System (MACS) for facilitating believable resource gathering behaviors on naviga-

tion graph in real-time strategy games. In the proposed system, the harvesters (i.e., ants) leave their base location (i.e.,

the ant colony nest), seek resources and drop pheromone on the way back to the nest. By following pheromone trails,

the harvesters are assured a path to resources and back home. The more ants that travel the path with resources found,

the stronger the pheromone grows1. While this path-finding search framework is based on Ant Colony Optimization

145 Xianshun Chen et al. / Procedia Computer Science 24 ( 2013 ) 143 – 151

(ACO) and similar to that in1, our proposed approach differs in three core manner. Firstly, in contrast to1, our path-

finding task focuses on navigation graphs, which is the typical search space representation for path-finding in digital

games8,9. Secondly, an individual learning phase in the form of a partial A-Star search is carried out by the harvester

whenever resources is located, which enhance the colony in converging to the best path-finding solutions more effi-

ciently. Third, memetic computation10, specifically memetic automaton and the process of imitation, are integrated

with the ant colony learning mechanism to form a meme-centric framework that adapts agents’ behaviors according to

the dynamics of the navigation graph. This effectively eliminates the need for tedious parameter fine-tuning in ACO

search framework.

Memetic computation is a paradigm that uses the notion of meme(s) as units of information encoded in compu-

tational representations for the purpose of problem-solving10,11,12. Meme(s) on its own is perceived as a form of

structured knowledge, for example, in the form of recurrent patterns. The proposed MACS is a meme-centric Ant

Colony System (ACS) framework in which the harvester’s path-finding and resource gathering knowledge captured

are extracted and represented as memes, which are internally encoded as the state transition rules (internal logic or

memotype) and externally expressed as ant pheromone on the graph edge (behavior or sociotype). The memotypesexist within the mind universe of a harvester and serves as the internal logics used by the harvester in selecting the

next edge to travel when it reaches a particular junction of the navigation map. On the other hand, the sociotypesare

expressed as the resultant pheromone deposited on the traveled edges by the harvester. These scents are then picked

up by other harvesters subsequently reaching the same location, which may then imitate the previous harvester’s be-havior in selecting the next edge. While the sociotypes are being evolved by pheromone evaporation and deposition

process within the Ant Colony System, the memetic evolution takes place at the same time within the harvesters’s

mind universe, whereby memotypes undergo selection, transmission and variation. Through the inter-play between

the memetic evolution and ant colony, harvester bots or more precisely memetic automaton, are able to acquire in-

creasing level of capability in exploring complex dynamic game environment and gathering resources in an adaptive

manner, producing consistent and impressive resource gathering behaviors11,13.

The rest of the paper is organized as follows: Section 2 begins with the detailed description of the RTS resource

gathering problem to be solved in which harvesters are only equipped with the knowledge of its immediate sur-

rounding. Subsequently, Section 3 discusses the design of the Memetic Ant Colony System, focusing on the core

contributions towards facilitating believable resource gathering behaviours in RTS games, namely, memetic represen-

tation, ACS and individual learning, as well as memetic evolution. In Section 4, the empirical assessment of MACS is

considered. Empirical results obtained show that MACS producing consistent and impressive resource gathering be-

haviors even when the harvesters are only given the limited knowledge of its immediate surroundings during resource

gathering. Finally, the last section ends with a brief conclusion of the presented research.

2. RTS Resource Gathering Problem

Resource gathering is arguably one of the most fundamental AI task in RTS games. In the resource gathering

problem we defined here, the harvester is only given the knowledge of its immediate surrounding. Suppose that the

navigation graph is denoted as G = (V, E), which consists of a set of vertices, V = {vi} and a set of edges E = {ei j} in

which an edge ei j = (vi, v j) connect vertices vi and v j. We denote Gk(t) ⊂ G as the partial graph visible to harvester

agt(k) at time step t, and we further denote Ei = {ei j|ei j = (vi, v j), ei j ∈ E, v j ∈ V} and Vi = {v j|v j ∈ ei j, ei j ∈ Ei}as the set of edges and vertices connecting to the vertex vi. When harvester agt(k) reaches vertex vi at time step t,its visibility only extends up to Vi, Ei, i.e., Gk(t) = (Vi, Ei). With such a limited visibility, the harvester agt(k) must

gather knowledge about the graph structure of the navigation map on which it explores by sharing information and

cooperating with other harvesters in the game environment.

A harvester is only given the knowledge of its immediate surroundings in resource gathering. This constraint is

important from the practical point of views in the context of RTS games. Firstly, modern game environments are

becoming frequently nondeterministic with imperfect knowledge and large numbers of quite dynamic variables14,15,

e.g. the obstacles may change location such as sudden destruction of a bridge, and the resources and navigation mesh

may also change dynamically. As a result, the creation of an AI resource gathering subsystem must be designed with

minimum use of the preprocessed information (e.g. the full visibility of the entire navigation graph) in order to handle

a fully dynamic navigation graph that deal with fast changes. Secondly, the dynamics of the world should also take

146 Xianshun Chen et al. / Procedia Computer Science 24 ( 2013 ) 143 – 151

into account that the game agents can change the states of the world, e.g., a resource location may become depleted

due to harvesters gathering resources from it, or soldier units causing the terrain to change after bombardment on a

piece of land. Moreover, a game scene can have many thousands of moving objects at peak times15, the resource

gathering behaviors in RTS games have to frequently handle the vast number of game agents as well as the time syn-

chronization among them in a persistent game world which can undergo large-scale physical change affecting dozen

to hundreds of characters. In addition, modern RTS usually features large-scale navigation graph that may undergo

large physical change as well. Consequently, from the consideration of computational efficiency, it is important to

minimize the amount of information to be processed by a single harvester during resource gathering, which can be

done by processing only the local navigation knowledge. Thirdly, the desired illusion of intelligence manifests best

when game agents emulate the behavior of a human player. That is, believable resource gathering behaviors of the

harvester should be imperfect to a degree, being either fun and beatable or perhaps simply irrational, as human beings

often are. The seemingly simple aspect of seeking and returning resources takes on new meaning in RTS games when

now the computer controlled players have to play on equal grounds as a human player, without cheating and equipped

with the same amount of incomplete map information.

3. Memetic Ant Colony System

In this section, we introduce the proposed memetic ant colony system for RTS resource gathering problem. In each

generation g, a colony of ants (i.e., harvesters) is generated, in which each harvester agt(k) carries a path memory Gk

and a tabu list tl(k). The path memory Gk records the edges and vertices that agt(k) traveled up to the current time

step and the tabu list tl(k) records the edges recently traveled by agt(k). agt(k) also carries its memotype M(k), which

encodes a set of state transition rules that determine which edge ei j to travel next when it reaches vi.

The harvesters start by moving out of their base location v0 (i.e., the ant colony nest) in search of resources. As

each harvester agt(k) has only a limited sight range Gk(t) = (Vi, Ei) when it reaches vi, the harvester must make use of

its available memotype M(k) to analyze the sociotype memes available in its surroundings. These sociotype memes

have been deposited as pheromone on the edges Ei by other harvesters previously passing over vi. agt(k) then imitates

the behavior of the harvesters who have successfully found a resource location or explore as far as possible to search

for resources. After agt(k) passes over ei j, it expresses a sociotype meme into the pheromone trail τi j on ei j and update

its tabu memory tl(k) to include ei j, so that harvesters can avoid potentially dangerous paths or paths that have been

recently traveled by other harvesters or itself, based on the pheromone that is laid on the edges as well as its tabu

memory tl(k).

After a harvester agt(k) successfully finds a resource location vI , the fitness of M(k) is increased by 1, i.e.,

f itness(k) = f itness(k) + 1. Subsequently, an individual learning is performed in which agt(k) pathfinds the shortest

path Pbest(v0, vI) from v0 to vI using partial A-Star search on its path memory Gk, while sociotype memes are then

deposited as pheromone on the edge set {ei j|ei j ∈ Pbest(v0, vI)}. Periodically, the pheromone on the edges traveled

by all harvesters are evaporated to encourage the harvesters in selecting paths that will lead them to already found

resource locations as well as to encourage them in exploring edges that have not yet being uncovered, in the hope of

finding new additional resources.

At the same time, when harvesters perform the ACS-like resource gathering activity, harvesters self-configure their

memotype encoded transition rules by subjecting them in a memetic evolutionary cycle, where memotype undergo

meme selection, transmission and variation in the harvester’s mind universe according their fitness.

The MACS search scheme include three key aspects: i) The memotype and sociotype representations of memes in

the MACS, ii) the ACS and A-Star individual learning mechanisms that promote the imitation of sociotype memes

among harvesters, and iii) the memetic evolution mechanisms that govern the evolution of memotype encoded state

transition rules.

3.1. Memetic Representation

In memetic computation, the memotype representation refers to knowledge or memes internal to the mind universal

of an harvester, while sociotype refers to the social expression of a meme. In the proposed MACS, sociotype is

expressed in the form of pheromone τ(r, s) that is deposited on the edge (r, s). On the other hand, the memotype

147 Xianshun Chen et al. / Procedia Computer Science 24 ( 2013 ) 143 – 151

representation of memes encode the state transition rules to provide a direct way in balancing between exploration

of new edges and exploitation of a priori and accumulated knowledge about the path-finding to known resource

locations. Structurally, the memotype representation M(k) is a continuous-value string that encodes the following

state transition rules:

M(k) = Tk ρ lk Q0 amin bmin

where Tk is the tabu tenure (i.e. the max length) of the tabu list tl(k) carried by harvester agt(k), rho is the local

pheromone update parameter for agt(k) to express its sociotype memes into the pheromone trail on the edge it just

traveled. lk is the lifetime of agt(k), Q0 is a parameter that balances between exploitation and exploration in the state

transition rules. amin and bmin are inhibit parameters that discourage harvesters from traveling to vertices that have

been recently traveled by the harvester or potentially dangerous (e.g., obstacles ahead).

The memotype and sociotype memes in MACS work in a complementary nature. Memotype are internal logic

blocks of an harvester when making decision on path-finding, while the sociotype are behaviors expressed by har-

vesters which are transmitted to other harvesters via pheromone. When reaching a vertex r, harvester agt(k) may

decide to imitate a particular successful harvester agt(l) previously passing this vertex, and select the edge agt(l) has

traveled. The behavior imitation is accomplished via the memotype state transition rules and sociotype pheromone

information as follows: a harvester agt(k), whose current visited vertex is r, chooses the next vertex s from its sur-

rounding vertices V(r) = {μ|(r, μ) ∈ E} to move by applying a set of state transition rules. If V(r) contains a resource

location vI , then s = vI . Otherwise s is selected from V(r) using the pseudo-random-proportional rule:

s ={

arg maxμ∈V(r) P(μ) if p ≤ Q0;

S if p > Q0.(1)

In Eqn. 1, P(μ) denote the preference for selecting vertex μ ∈ V(r), p is a random number that is uniformly

distributed in (0, 1] and Q0 is taken from M(k). If p ≤ Q0, the pseudo-random-proportional rule favors exploitation (sis chosen as the vertex such that edge (s, r) is the edge on which the preference P(μ = s) is the highest). If p > Q0,

the algorithm favors exploration and s is assigned a random variable S selected with the probability P(S , r):

P(S , r) =P(S )∑

u∈V(r) P(u)(2)

Eqn. 2 is known as the random-proportional rule in that it favors transitions toward vertices having higher prefer-

ence P(μ).For each neighboring vertex μ ∈ V(r), its preference P(μ) is given by:

P(μ) =

{τ(r, μ) μ � tl(k)

τmin ∗ amin μ ∈ tl(k)(3)

where τ(r, μ) is the pheromone on the edge (r, μ), and τmin is the minimum pheromone value. 0 ≤ amin ≤ 1 is from

M(k). Eqn. 3 discourages the harvester agt(k) from selecting a vertex that is in its tabu list tl(k).

The state transition rules in Eqns. 1, 2, 3 for resource gathering differ from that of the ACS state transition rules

in traditional TSP path planning in that there is a probability that the harvester agt(k) will revisit a vertex which it

has visited in the past, so as to explore its neighboring vertices again in hope of discovering a resource location.

Furthermore, it does not assume a fully connected graph as in TSP path planning.

Initially, the pheromone level of any edge not traveled by any harvester is assumed to have a pheromone level of

τ0. After agt(k) travels the edge (r, s), it expresses its temporal presence at the edge (r, s) via pheromone update as

given by:

τ(r, s)← (1 − ρ) · τ(r, s) + ρ · τ0 (4)

where the local decay parameter 0 < ρ < 1 is taken from the memotype M(k) of harvester agt(k). If agt(k) dies as

a result of an ambush at vertex s, it also communicates this acknowledgement on edge (r, s) via the pheromone update

as given by:

τ(r, s)← τmin ∗ bmin (5)

148 Xianshun Chen et al. / Procedia Computer Science 24 ( 2013 ) 143 – 151

where the pheromone inhibit parameter 0 < bmin < 1 is taken from the memotype M(k) of harvester agt(k).

Eqns. 4 and 5 allow harvester to share knowledge via sociotype meme expression in terms of pheromone changes

on the edge. Thus subsequent harvesters can avoid paths potentially dangerous, or recently traveled by other harvesters

or by itself.

3.2. ACS and Individual Learning

Here we detail the Ant Colony System (ACS) process of pheromone evaporation and deposition as well as the

individual learning involved in the resource gathering. After an harvester agt(k) successfully finds a resource location

vI , the path-finding knowledge or sociotype meme can be transmitted to other harvesters via the global pheromone

update, in which agt(k) raises the pheromone level on the traversed edges stored in its path memory Gk. However,

agt(k) may have to take many unnecessary twist and turns during its search for the resource location. Thus it is

important to locally refine its path memory Gk. The local refinement is performed through an individual learning

stage in which agt(k) path finds the shortest path Pbest(v0, vI) from v0 to vI using partial A-Star search on its path

memory Gk. The A-Star search is partial as it is only restricted to the partial graph represented by Gk. Sociotype

memes are then deposited as pheromone on the edge set {(r, s)|(r, s) ∈ Pbest(v0, vI)}, using the global update rule as

given by:

τ(r, s)← τmax∀(r, s) ∈ Pbest(v0, vI) (6)

where τmax is the maximum pheromone level.

Periodically, the pheromone on the edges traveled by all harvesters are evaporated, to encourage the harvesters in

selecting paths that will lead them to the found resource locations

τ(r, s)← (1 − α) ∗ τ(r, s),∀(r, s) ∈ Gk,∀k (7)

It is worth noting that Eqn 7 only applies to edges that have been traversed by the harvesters, while the pheromone

level of any untraversed edge is maintained at τ0. This is to encourage harvesters in exploring edges that have not

covered in the hope of finding more resources, particularly when the current resources found have been depleted or

when the previously found paths to the resource locations are blocked.

3.3. Memetic Evolution

Many of the parameters within the memotype memes of an harvester specify the behavior of the state transition

rules during path-finding. The role of memetic evolution is to auto-tune the behavior of the state transition rules so

that the harvester can self-adapt to the dynamics of the navigation graph. The memetic evolution inter-plays with

the ACS resource gathering described in Subsections 3.1 and 3.2. In the lifetime of an harvester agt(k), each time

it successfully finds a resource location, the fitness f itness(k) of its currently associated memotype meme M(k) is

increased by 1, i.e., f itness(k) = f itness(k) + 1. This gives rise to a selection pressures among the memotype memes

residing in the meme pool constituting memotypes memes being used by harvesters. In a memetic evolution cycle,

these memotype memes undergo meme selection, transmission and variation according their fitness f itness(k), which

give rise to the new generations of memotype memes. In our current implementation, a binary tournament selection

scheme is used to select the memotype memes either for path-finding or memetic reproduction. Redcliff Blending

is used for memetic transmission in which a new meme is produced by inheriting properties from its parent memes.

With respect to the memetic variation process, a normal distribution mutation is applied to heuristically modify the

memetic materials.

The memetic evolution allows harvesters to adapt their state transition rules to the dynamics of a game world which

can undergo fast and large-scale physical change affecting dozen to hundreds of game agents, by balancing between

exploitation and exploration on the navigation graph.

4. Empirical Study

In this section, we evaluate the proposed memetic ant colony system for resource gathering on a simulated RTS

environment, which is depicted in Fig. 1. In the simulated environment, the graph for resource gathering is randomly

149 Xianshun Chen et al. / Procedia Computer Science 24 ( 2013 ) 143 – 151

Obstacle

Start Point

Supply Point (Resource)

AntExplored Edge

Capacity Bar

Fig. 1. Memetic Ant for Resource Gathering.

Table 1. Summarization of parameter configurations in plentiful supply, medium supply and little supply.

Supply Point Obstacle Memetic AntsPlentiful Supply 20 5 20

Medium Supply 10 10 20

Little Supply 5 20 20

generated. We can configure the number of start point, obstacles, supply points (resource) and ants at the right plane

as shown in Fig. 1. Each supply point has a capacity as shown on its top, and the capacity bar will become empty

when the supply point has been exhausted. The edge with highlighted color denotes the path that has been explored

by the memetic ants. More importantly, the whole structure of the generated graph is non-familiar to the memetic

ants, and only the immediately surrounding information is available to the respective memetic ants.

Further, three scenarios of resource gathering, namely plentiful supply, medium supply and little supply are studied

in this experiment. In particular, for plentiful supply, extensive supply points are available to the memetic ants, and

little obstacles existed on the graphs. Medium supply has equal size of supply points and obstacles on the navigation

graph, while little supply holds lots of obstacles in the navigation graph but little supply points available to the memetic

ants. The detailed parameter configuration of each scenario is summarized in Table 1. Fig. 2 presents the navigation

graph and respective distribution of supply points and obstacles in each resource gathering scenario considered. Since

the number of memetic ants is kept fixed in all the three scenarios, the difficulty of resource gathering task increases

from plentiful to little supply.

All the simulated results of the proposed memetic ant colony system on plentiful supply, medium supply and little

supply resource gathering scenarios are presented in Fig. 3. As can be observed, the resources in both easy (i.e., Fig.

150 Xianshun Chen et al. / Procedia Computer Science 24 ( 2013 ) 143 – 151

(a) Plentiful Supply (b) Medium Supply (c) Little Supply

Fig. 2. Distributions of Supply of Plentiful Supply, Medium Supply and Little Supply Resource Gathering Scenarios.

3(a) plentiful supply) and difficult (i.e., Fig. 3(c) little supply) resource gathering have all been successfully found

by the memetic ants (i.e., Capacity = 0). The shortest paths from the start point to each supply point (resource) are

all highlighted in red. It is worth noting that the graphes of the three resource gathering scenarios considered are all

different, the consistent performance of the proposed MACS confirmed its effectiveness in path-finding when limited

information of the navigation graph is available to the harvesters.

5. Conclusion

This paper proposes the conceptual modeling of a memetic ant colony system (MACS) for facilitating believableresource gathering behaviours in RTS games, where harvesters are equipped with only limited knowledge about its

immediate surroundings. In the proposed MACS, the harvester’s path-finding and resource gathering knowledge

captured are extracted and represented as memes, which are internally encoded as state transition rules (memotype)

and externally expressed as ant pheromone on the graph edge (sociotype). Through the inter-play between the memetic

evolution and ant colony, harvesters as memetic automaton spawned from an ant colony are able to acquire increasing

levels of capabilities in exploring the complex dynamic game environment and gathering resources in an adaptive

manner. The empirical study shows that the proposed MACS is able to produce consistent and believable resource

gathering behaviors even when only limited knowledge of its immediate surroundings is available in the process of

resource gathering.

Acknowledgment

This research is supported by Multi-plAtform Game Innovation Centre (MAGIC), funded by the Singapore Na-

tional Research Foundation under its IDM Futures Funding Initiative and administered by the Interactive & Digital

Media Programme Office, Media Development Authority.

151 Xianshun Chen et al. / Procedia Computer Science 24 ( 2013 ) 143 – 151

Capacity = 0

Shortest Path

Capacity = 0

Shortest Path

Capacity = 0

Shortest Path

(a) Plentiful Supply (b) Medium Supply (c) Little Supply

Fig. 3. Paths found by The Proposed Memetic Ant Colony System in Plentiful Supply, Medium Supply and Little Supply Resource Gathering

Scenarios.

References

1. J. Dunn, Ant colony organization for mmorpg and rts creature resource gathering, AI Game Programming Wisdom 3 (2005) 495–505.

2. I. Umarov, M. Mozgovoy, Believable and effective ai agents in virtual worlds: Current state and future perspectives, International Journal ofGaming and Computer-Mediated Simulations (IJGCMS) 4 (2) (2012) 37–59.

3. T. S. Standley, R. E. Korf, Complete algorithms for cooperative pathfinding problems, in: IJCAI 2011, pp. 668–673.

4. D. Silver, Cooperative pathfinding, AI Game Programming Wisdom 3 (2005) 99–111.

5. R. Jansen, N. Sturtevant, A new approach to cooperative pathfinding, in: Proceedings of the 7th international joint conference on Autonomousagents and multiagent systems - Volume 3, AAMAS ’08, International Foundation for Autonomous Agents and Multiagent Systems, Richland,

SC, 2008, pp. 1401–1404.

6. M. R. Jansen, N. R. Sturtevant, Direction maps for cooperative pathfinding., in: AIIDE’08, 2008, pp. –1–1.

7. G. Snook, Simplified 3d movement and pathfinding using navigation meshes, Game Programming Gems, (2000) 188–204.

8. C. McAnlis, J. Stewart, ntrinsic detail in navigation mesh generation, Game Programming Wisdom 4 (2008) 95–112.

9. D. Hamm, Navigation mesh generation: An empirical approach, Game Programming Wisdom 4 (2008) 113–123.

10. X. Chen, Y.-S. Ong, M.-H. Lim, K. C. Tan, A multi-facet survey on memetic computation, IEEE Trans. Evolutionary Computation 15 (5)

(2011) 591–607.

11. Y. S. Ong, M. H. Lim, X. S. Chen, Research frontier: Memetic computation - past, present & future, IEEE Computational IntelligenceMagazine 5 (2) (2010) 24–36.

12. L. Feng, Y. S. Ong, I. W. H. Tsang, A. H. Tan, An evolutionary search paradigm that learns with past experiences, IEEE World Congress onComputational Intelligence, Congress on Evolutionary Computation, 1 – 8.

13. L. Feng, Y. S. Ong, A. H. Tan, X. S. Chen, Towards human-like social multi-agents with memetic automaton, IEEE Congress on EvolutionaryComputation (2011) 1092–1099.

14. P. Marden, F. Smith, Dynamically updating a navigation mesh via efficient polygon subdivision, Game Programming Wisdom 4 (2008) 83–94.

15. R. Axelrod, Navigation graph generation in highly dynamic worlds, AI Game Programming Wisdom 4 (2008) 125–141.